Convolutional 2D Knowledge Graph Embeddings resources. Source code coming soon.
Paper: 2D Convolutional Graph Embeddings
Used in the paper, but do not use these datasets for your research: FB15k and WN18
This repo supports Linux and Python installation via Anaconda.
- Install PyTorch using Anaconda
- Install the requirements
pip install -r requirements - Run the preprocessing script for WN18RR, FB15k-237, and YAGO3-10:
sh preprocess.sh - (Optional) Install spacy English models with
python -m spacy download en. - You can now run the model
This is a bit messy, but it works.
- Select your model and dataset in the main.py script
- You can pass common parameters like learning rate (lr), dropout values (input_drop, hidden_drop, feat_drop) as parameters to the script like so:
CUDA_VISIBLE_DEVICES=0 python main.py input_drop 0.2 hidden_drop 0.3 feat_drop 0.2 lr 0.003
The command above in conjunction with ConvE as a model will yield state-of-the-art results for most link prediction datasets.
For the reverse model, you can run the provided file with the name of the dataset name and a threshold probability:
python reverse_rule.py WN18RR 0.9
To run it on new datasets, either adjust the path structure in the file, or copy your dataset folder into the data folder and make sure your dataset split files have the name train.txt, valid.txt, and test.txt which contain tab separated triples of a knowledge graph.